Multi-objective generators in deep learning

Abstract

Machine-learning data generators use an additional objective to avoid generating data that is too similar to any previously known data example. This prevents plagiarism or simple copying of existing data examples, enhancing the ability of a generator to usefully generate novel data. A formulation of generative adversarial network (GAN) learning as the mixed strategy minimax solution of a zero-sum game solves the convergence and stability problem of GANs learning, without suffering mode collapse.

What is claimed is: 
     
       1. A computer-implemented method comprising:
 training, with a computer system that comprises a set of processor cores, a machine-learning real vs. generated discriminator to discriminate whether output from a data generator is real or generated; 
 training, with the computer system, through machine learning, a multi-category classifier such that a target output of the multi-category classifier comprises a target activation level for each of the multiple classification categories; and 
 training, with the computer system, through machine learning, a machine-learning data generator, wherein training the machine-learning data generator comprises training the machine-learning data generator through machine-learning with multiple joint objectives, wherein the multiple joint objectives comprise:
 a first objective that the machine-learning data generator generate, as output patterns, data patterns that the machine-learning real vs, generated discriminator system classifies as real; 
 a second objective that is different from the first objective, wherein the second objective imposes a penalty on a cost function for the data generator for an output pattern that is within a threshold distance of a nearest neighbor candidate pattern in a set of nearest neighbor candidate patterns for the output pattern; and 
 a third objective that is different from the first and second objectives, wherein the third objective comprises a multi-target objective from the multi-category classifier, such that output patterns from the machine-learning data generator do not match the training data for the multi-category classifier. 
 
 
     
     
       2. The computer-implemented method of  claim 1 , wherein:
 the machine-learning data generator comprises a feed-forward neural network; and 
 training the machine-learning data generator comprises training the machine-learning data generator using stochastic gradient descent. 
 
     
     
       3. The computer-implemented method of  claim 1 , wherein the machine-learning data generator comprises a stochastic autoencoder. 
     
     
       4. The computer-implemented method of  claim 3 , wherein the stochastic autoencoder comprises a variational autoencoder. 
     
     
       5. The computer-implemented method of  claim 3 , wherein the stochastic autoencoder comprises a stochastic categorical autoencoder network. 
     
     
       6. The computer-implemented method of claim of  claim 1 , wherein the machine-learning data generator comprises a generative adversarial network. 
     
     
       7. The computer-implemented method of  claim 1 , wherein the second objective back propagates partial derivatives of the penalty through the machine-learning data generator when an output pattern from the machine-learning data generator is within the threshold distance of the nearest neighbor candidate pattern in the set of nearest neighbor candidate patterns for the output pattern. 
     
     
       8. The computer-implemented method of  claim 7 , further comprising generating, by the computer system, the set of nearest neighbor candidate patterns with a Hopfield network. 
     
     
       9. The computer-implemented method of  claim 7 , further comprising generating, by the computer system, the set of nearest neighbor candidate patterns with an auto-associative memory. 
     
     
       10. The computer-implemented method of  claim 7 , further comprising computing, by the computer system, a distance between the output pattern and the nearest neighbor candidate pattern. 
     
     
       11. The computer-implemented method of  claim 10 , wherein computing the distance comprises computing a Euclidean distance between the output pattern and the nearest neighbor candidate pattern. 
     
     
       12. The computer-implemented method of  claim 10 , wherein the penalty is inversely related to the distance between the output pattern and the nearest neighbor candidate pattern. 
     
     
       13. The computer-implemented method of  claim 1 , wherein the first objective imposes a generated-data penalty term that is back-propagated to the machine-learning data generator when an output pattern from the machine-learning data generator is determined by the machine-learning real vs. generated discriminator system to be generated rather than real. 
     
     
       14. The computer-implemented method of  claim 1 , wherein the multiple joint objectives comprise a fourth objective that is an output of a machine-learning distortion detector, such that the fourth objective back propagates an excess distortion penalty term to the machine-learning data generator. 
     
     
       15. The method of  claim 1 , wherein:
 training the real vs. generated discriminator comprises training the real vs. generated discriminator with real vs. generated training data; and 
 training the multi-category classifier comprises training the multi-category classifier with training data from the real vs. generated training data. 
 
     
     
       16. A machine-learning computer system comprising:
 a set of processor cores; and 
 computer memory that is in communication with the set of processor cores, wherein the computer memory stores software that when executed by the set of processor cores, causes the set of processor cores to:
 train a machine-learning real vs. generated discriminator system that is trained, with real data in a training set of real data, to discriminate real data from generated data; 
 train through machine learning a multi-category classifier such that a target output of the multi-category classifier comprises a target activation level for each of multiple classification categories; 
 train a machine-learning data generator with multiple joint objectives, wherein the multiple joint objectives comprise:
 a first objective that the machine-learning data generator generate, as output patterns, data patterns that the machine-learning real vs. generated discriminator system classifies as real; 
 a second objective that is different from the first objective, wherein the second objective imposes a penalty on a cost function for the data generator for an output pattern that is within a threshold distance of a nearest neighbor candidate pattern in a set of nearest neighbor candidate patterns for the output patter; and 
 a third objective that is different from the first and second objectives, wherein the third objective comprises a multi-target objective from the multi-category classifier, such that output patterns from the data generator do not match the training data for the multi-category classifier. 
 
 
 
     
     
       17. The machine-learning computer system of  claim 16 , wherein:
 the machine-learning data generator comprises a feed-forward neural network; and 
 training the machine-learning data generator comprises training the machine-learning data generator using stochastic gradient descent. 
 
     
     
       18. The machine-learning computer system of  claim 16 , wherein the machine-learning data generator comprises a stochastic autoencoder. 
     
     
       19. The machine-learning computer system of  claim 18 , wherein the stochastic autoencoder comprises a variational autoencoder. 
     
     
       20. The machine-learning computer system of  claim 18 , wherein the stochastic autoencoder comprises a stochastic categorical autoencoder network. 
     
     
       21. The machine-learning computer system of  claim 16 , wherein the machine-learning data generator comprises a generative adversarial network. 
     
     
       22. The machine-learning computer system of  claim 16 , wherein the second objective back propagates partial derivatives of the penalty through the machine-learning data generator when an output pattern from the machine-learning data generator is within the threshold distance of the nearest neighbor candidate pattern in the set of nearest neighbor candidate patterns to the output pattern. 
     
     
       23. The machine-learning computer system of  claim 22 , wherein the computer memory stores software that when executed by the set of processor cores causes the set of processor cores to generate the set of nearest neighbor candidate patterns with a Hopfield network. 
     
     
       24. The machine-learning computer system of  claim 22 , wherein the computer memory stores software that when executed by the set of processor cores causes the set of processor cores to generate the set of nearest neighbor candidate patterns with an auto-associative memory. 
     
     
       25. The machine-learning computer system of  claim 16 , wherein the first objective imposes a generated-data penalty term is back-propagated to the machine-learning data generator when an output pattern from the machine-learning data generator is determined by the machine-learning real vs. generated discriminator system to be generated rather than real. 
     
     
       26. The machine-learning computer system of  claim 16 , wherein the multiple joint objectives comprise a fourth objective that is an output of a machine-learning distortion detector, such that the fourth objective back propagates an excess distortion penalty term to the machine-learning data generator. 
     
     
       27. The machine-learning computer system of  claim 16  wherein the computer memory stores software that when executed by the set of processor cores causes the set of processor cores to:
 train the real vs. generated discriminator training the real vs. generated discriminator with real vs. generated training data; and 
 train the multi-category classifier training the multi-category classifier with training data from the real vs. generated training data.